The job is not choosing models. An AI consultant’s work is choosing which of your processes to change, getting one of them running inside the tools you already pay for, and producing the before-and-after number that shows it moved. Everything else is packaging.
That is an unglamorous description, and it is the one worth paying for. The Bureau of Labor Statistics does not track "AI consultant" as an occupation, so there is no official wage, no official headcount, and no rate card. What follows is what the work actually consists of, what the closest measured occupation pays, and the five questions to ask before you sign anything.
What does an AI consultant actually do in the first few weeks?
Inventories the processes that touch money, then picks one. The deliverable of week one is not a model or a tool; it is a short list of where a lead enters, who touches it, how long each handoff takes, and which of those steps you can measure today.
That list is harder to produce than it sounds, because most of it is not written down anywhere. The phone gets answered by whoever is closest to it. A quote goes out when someone has an hour. Follow-up happens if the person who took the call remembers. None of that shows up in a software audit, and all of it shows up in the revenue.
So the first few weeks look like interviews and stopwatches more than configuration. Four processes carry almost all of the exposure in a service business: the inbound call, the quote, the follow-up, and the collection. For each one the question is the same, and it is deliberately boring. How does a job get from the previous step to this one, how long does that take, and what happens when the person responsible is busy?
Anyone who opens with a tool recommendation skipped the only part that is hard. The order of the work, phase by phase, is in what we actually do in the first 90 days; this post is about what the work is, not the calendar it runs on.
What does this kind of work cost on the open market?
There is no BLS occupation called "AI consultant", so the honest anchor is the closest one it does track: management analysts, with a median annual wage of $101,860 in May 2025, about $48.97 an hour. The band around that median is wide. The bottom 10% of that occupation earned less than $60,640 and the top 10% more than $171,640.
Two things about that number before anyone quotes it back at you. It is not what AI consultants earn, because no one measures what AI consultants earn; it is what the nearest measured occupation earns, and the gap between the two is unknown. And the May 2025 estimates were collected during the federal funding lapse that ran from October to November 2025, which the BLS says delayed collection without changing the methodology.
Used properly, that band is a sanity check rather than a price. A proposal, a salary, or a retainer all have to resolve to someone’s time at some rate, and the rate for experienced process work in the United States sits inside those two numbers. A quote far below the bottom of the band is buying someone junior or someone offshore, which can be the right call as long as you know that is the trade. A quote far above the top is buying something other than the analysis, and it is fair to ask what.
Why does a $5M company hire this and a $500K one does not?
Because at a few million in revenue the bottleneck stops being effort and starts being handoffs. There is finally a process that runs often enough to be worth instrumenting, and enough people touching it that nobody can hold the whole thing in their head.
Below that, the owner usually is the process. The quote goes out fast because the person who priced it answered the phone. There is not much to coordinate, so coordination software buys very little, and the return on a consultant is close to zero. Above it, the same volume that used to be a good problem starts arriving faster than the handoffs can absorb it: the call that rings out at 4:50pm, the quote that sits in a draft folder overnight, the job that nobody followed up on because the person who owned it was on a roof.
For scale of where the market sits, a national survey of nearly 1,000 U.S. small businesses with annual revenue between $25,000 and $5 million (Reimagine Main Street with PayPal, fielded May 2025) sorted respondents by how far they had gone with AI. The largest group, 51%, were "Explorers": experimenting without having committed to anything.
One caveat about reading yourself into that number. The survey’s revenue range ends at $5 million, so a company at $5 million sits at the top edge of the sample rather than in the middle of it. The 51% describes businesses up to $5 million in revenue, and it is useful for seeing where your peers are, not for deciding that you are late. Most of that group is experimenting because experimenting is cheap and committing is not.
What should the engagement hand you that you can actually audit?
One process running inside the tools you already pay for, plus the before-and-after number for that process. If the handoff at the end of the engagement is a document, you bought a document.
The gap this closes is specific, and one vertical has it measured. Among 1,018 roofing companies ServiceTitan surveyed in late 2025, only 4% had AI features built directly into their CRM, while 25% used an external LLM tool. ServiceTitan sells the CRM those contractors would be integrating into, so read that as the vendor’s own read of its market rather than a neutral census, and note that it describes roofing contractors specifically, not small businesses generally.
Even with those caveats, the shape of the finding is the useful part. A chatbot in another tab is not integration; it is a person copying and pasting. It helps whoever is sitting in front of it and disappears the moment that person is on a call, which is exactly when the work was failing before. Integrated means the system is in the path the job already travels: it answers the line, it writes to the record, it sends the follow-up whether or not anyone remembered.
So the thing to audit at the end is mechanical. Open the tool your team already uses every day and find the new behavior inside it. If you have to open something else, or ask someone to run something for you, the process did not change; a capability got added next to it.
What is not the job?
Not a tool tour, not a strategy deck, and not a pilot that never ends. Those three failure modes have one thing in common: none of them changes a step that a job passes through, so none of them can produce a number at the end.
It helps to know what owners themselves say stops them, because it is almost never the technology. In that 2025 survey the three reasons given for standing still were data privacy and security (38%), no time or resources to explore the tools (37%), and no clear use case or return yet (34%). These are reasons owners gave, not causes anyone measured, and respondents could pick more than one, so the shares do not add to 100.
Reimagine Main Street with PayPal, 2025 · multiple answers allowed, so the shares do not add to 100
View data
| Item | Value |
|---|---|
| Data privacy and security | 38% |
| No time or resources to explore the tools | 37% |
| No clear use case or return yet | 34% |
The same survey asked what would move them. Clearer evidence of return came first at 74%, tools that are easier to use came second at 73%, and hands-on training was the most-requested form of support across every segment. Those are the owners' words about what they want, not a promise about what any engagement delivers.
Read the two lists together and the test for a consultant writes itself. A deck does not reduce privacy exposure, does not give anyone time back, and does not produce return evidence. A pilot with no end date actively makes the third problem worse, because it consumes the attention that was the second-biggest constraint and never arrives at a number. If the work does not visibly reduce one of those five lines, the deliverable is the deliverable.
How do you know whether it paid for itself?
Pick the number before you start, and make sure it is one you can read without asking anyone. Calls answered, hours from form submit to first reply, quotes followed up, no-shows. Then compare the four weeks before to the four weeks after, with the same definition on both sides.
The "same definition" clause is where most of these measurements fall apart. If "answered" meant a human picked up in the before period and includes a voicemail transcript in the after period, the improvement is in the definition. Write down what counts on day zero, in one sentence, and do not revise it once you can see which way the number is going.
This is also the line between experimenting and integrating, which is the whole subject of how many small businesses actually use AI: experimenting produces a yes on a survey, and integrating produces a change you can find in the books.
What should you ask before signing anything?
Five questions, and the answers should be specific enough to be wrong.
Which process are you touching first, and why that one instead of the other three? What number gets measured before and after, and who reads it? Which of our tools is this running inside when you leave? What happens on the day it fails, and who gets the call? And which part of this stays a person’s job on purpose?
An engagement that can answer all five in plain sentences is one you can hold to account in ninety days. If the answers arrive as categories rather than specifics, the scoping work has not been done yet, and the first invoice would be paying for it twice.
Questions owners ask
What does an AI consultant actually do?
Picks which of your processes to change, gets one running inside the tools you already use, and measures the before and after. The first deliverable should be an inventory of the steps where a lead or a job changes hands, not a tool recommendation. If the engagement opens with software, the hard part was skipped.
How much does an AI consultant cost?
There is no published rate card for the category, and anyone quoting an industry average is guessing. The useful anchor is the closest occupation the Bureau of Labor Statistics tracks, management analysts, with a median annual wage of $101,860 in May 2025, the bottom 10% under $60,640 and the top 10% over $171,640. Any proposal, hire or retainer sits somewhere inside that band.
Does a company with $5 million in revenue need an AI consultant?
It needs one process fixed and measured, which is sometimes a consultant and sometimes an afternoon with the software you already pay for. In a national survey of nearly 1,000 U.S. small businesses with annual revenue between $25,000 and $5 million (Reimagine Main Street with PayPal, fielded May 2025), the largest group, 51%, were experimenting with AI without having committed to anything. The question is not whether to hire someone; it is which single handoff is costing you the most.
Why do most small-business AI projects stall?
Because the reasons owners give are rarely technical. In that 2025 survey they named data privacy and security (38%), no time or resources to explore the tools (37%), and no clear use case or return yet (34%), while 74% said clearer evidence of ROI would move them. A project scoped to answer one of those is a project that finishes.
Is using ChatGPT the same as having AI in the business?
Not the same thing, and the gap is measurable. Among 1,018 roofing companies ServiceTitan surveyed in late 2025, 25% used an external LLM tool while only 4% had AI features built directly into their CRM. ServiceTitan sells the CRM, so read it as the vendor’s own take on its market. A chatbot in another tab still needs a person to copy the answer across.
Sources
- U.S. Bureau of Labor Statistics, Occupational Outlook Handbook: Management Analysts (national OEWS wage estimates, reference period May 2025) (2025)
- Reimagine Main Street with PayPal, national survey of nearly 1,000 U.S. small businesses with annual revenue between $25,000 and $5 million (fielded May 2025) (2025)
- Public Private Strategies Institute, AI on Main Street: What Small Businesses Really Think (same survey) (2025)
- ServiceTitan, 2026 Roofing & Exteriors Market Report (1,018 roofing companies, surveyed late 2025) (2026)
The first meeting starts with findings, not promises.The free Business Health Report reads your business and shows you what it found, in days.
Get Your Free Business Health Report